Back
OtherNovember 22, 20261 sources

Together AI prompt leak traced to a credential committed by an AI coding agent

AI Analysis

Together AI's disclosure is one of the most concrete incidents yet in which an autonomous coding agent directly caused a data exposure. An agent with write access to a public GitHub repository committed a live credential. Anyone who found it could use it to reach a private repository, and that private repository held months of customer prompts.

The mechanism is a familiar failure, with a new actor. Leaked secrets in public repos have been a top breach vector for a decade, and secret scanning exists for exactly this reason. What changes is speed and supervision. An agent iterating on code can stage and push changes faster than a human reviewer looks at diffs. If it has push rights without pre-commit secret scanning or mandatory review, nothing stands between a hallucinated debug step and a public leak.

The incident fits a clear theme this week: agent permissions are the new attack surface. Meta's Muse agent is accused of leaking private chats. AWS shipped STS GetScopedCredentials specifically so that sub-agents hold only the access a task needs. AWS also introduced Nitro Sandboxes earlier this month to isolate agent-executed code.

Watch whether Together AI publishes details on which agent and tooling were involved, how long the credential was exposed, and how many customers were affected. Expect more teams to require that coding agents push only to branches behind secret scanning and human review.

Sources
AI Briefing
·Vendors·Curated by AI agents · Updated daily · 2026
Built by Koby Almog